Pu-Huai Chen
SAR IMAGE GEOCODING USING A STEREO-SAR DEM AND AUTOMATICALLY
GENERATED GCPs
Pu-Huai CHEN, Ian DOWMAN,
University College London, U.K.
Department of Geomatic Engineering
idowman@ge.ucl.ac.uk
KEY WORDS: Automation, Control strategies, DEM, Geo-referencing, RADARSAT, Simulation, Spatial data, Stereo
SAR.
ABSTRACT
SAR image geocoding needs an accurate DEM, but such a DEM can be difficult to obtain. This paper shows that the
needs of geocoding can be satisfied using the DEM generated by employing a rigorous geometric model and an
optimized least squares correlation method with a region-growing approach. The geometric model used is independent
of any manually selected GCPs and is robust to orbit errors. The required number of GCPs for correcting any systematic
effects after space intersection is only two. When good quality orbit information is unavailable, GCPs or tie points are
essential for SAR image geocoding, but selection and measurement of these points is often difficult.
This paper proposes a SAR simulation technique to automatically provide control for geocoding without manually
selected GCPs or tie points. The GCPs were generated using a small known DEM chip of size 1km by 1km with
significant terrain relief. The RMS errors of the four automatically generated GCPs are less than 3.4m in range direction
and 4.6m in azimuth direction. RADARSAT SAR images have been tested using a reference DEM and an automatically
generated DEM to produce geocoded images. The RMS errors of the check points range from 13m to 21m in easting
and 16m to 21 in northing. The results derived in this paper demonstrate that the stereo-SAR generated DEM can be
used for geocoding in flat-moderate areas, and a higher level of automation can be achieved.
1 INTRODUCTION
Geocoding is to geometrically rectify a remotely sensed image, such as a Synthetic Aperture Radar (SAR) image,
according to a specific map projection and a Digital Elevation Model (DEM) to eliminate terrain induced image
distortions. The key factor to produce a geocoded radar image-map is the use of a correct DEM and reliable Ground
Control Points (GCPs). However, DEMs are not always available and may be difficult to produce, e.g. in cloud-covered
locations where optical data is not available or in underdeveloped areas where collecting GCPs is difficult. Two
distinctive and practical methods have been developed for generating a DEM from SAR data, including interferometric
SAR (IfSAR) and stereoscopic SAR, (Leberl, 1990).
In terms of frequent mapping and global monitoring, however, experiments of the repeat-pass space-borne IfSAR
system show that the method often gives poor results due to poor coherence and to different atmospheric and physical
conditions. Compared with IfSAR approach, the stereoscopic SAR method, based on measuring the co-ordinate
difference of a common ground point from an image pair and converting it to spatial data according to an appropriate
geometric model, does not require such tight conditions. Another advantage of the stereoscopic SAR method is the
ability of direct provision of the spatial information of any image point and of linear/area measures. It means that the
stereo SAR method is relatively flexible. In theory, there also remain other conditions for stereo SAR, such as a
reasonable intersection angle between each image for collecting distinctive parallaxes, allowing them to be transformed
into height data. This condition is less significant, since multiple options of incidence angle of space-borne SAR data
are available nowadays, such as RADARSAT data, (CSA, 1995). The Shuttle Radar Topography Mission (SRTM) was
launched in early 2000, it uses a single-pass IfSAR system and is designed to carry out terrain elevation mapping of
80% of the Earth’s land surface, (JPL, 2000). It will be shown later that the stereoscopic radargrammetric method can
be used as an alternative tool for general mapping to the IfSAR approach, whenever the IfSAR method is not applicable
or the IfSAR generated DEM is not available to common users.
A rigorous geometric algorithm for generating spatial information with an error model to be validated is necessary to
understand the full potential and limitations of extracting spatial information from stereoscopic SAR data. Also, the
requirements of human operations and GCPs have to be reduced. Early work on ERS-1 data at UCL suggests that the
stereoscopic radargrammetric approach is a promising tool for extracting elevation data from space-borne SAR data,
38
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B1. Amsterdam 2000.
Pu-Huai Chen
(Clark, 1991). Chen and Dowman (1996) proposed an analytic approach to carry out space intersection using the leastsquares adjustment for ERS-1 data and reported results without using any GCPs, if good quality orbit data is available.
Further consideration and refinement for the geometric model to be applied to the SAR data with inferior quality orbit
data, such as RADARSAT SAR imagery, has been carried out (Chen, 2000). Image correlation is one of the key steps
from analytic to digital radargrammetry. Pyramidal correlation strategy using the least-squares correlation method with
a region-growing approach has been proved useful to generate a parallax file from SIR-B SAR data, (Denos, 1992), and
a DEM can be generated from ERS-1 images (Twu, 1996). An optimized strategy for giving the pyramidal structure has
also been developed (Dowman and Chen, 1998).
Providing ground control is indeed a problem in digital radargrammetry. Manual operations for collecting GCPs from
radar images and maps are usually limited by visual image-map correlation, and are not always stable hence the GCP
quality cannot be assured. In order to cope with this problem, SAR image simulation techniques have been used for
provision of ground control in geocoding, for instance, Kwok et al. (1990) and Guindon (1995). The accuracy of the
automatically provided ground control relies on the quality of the simulator as well as on the outcome derived from the
real-simulated SAR image correlation. Poor quality simulated images may lead to poor results of real-simulated image
correlation. Previous work shows that most of the currently available SAR image simulators still require many human
interventions. A SAR image simulator is designed to provide ground control by using four corner elements of a known
DEM chip, according to a SAR imaging model to relate the DEM sensed and the SAR image , provided that a small
DEM chip with significant terrain relief is available (Chen, 2000).
This paper describes the use of the DEMs derived from stereo RADARSAT SAR data and the automatically derived
GCPs for geocoding to achieve a higher level of automation in radargrammetry. Section 2 reviews the algorithms
developed briefly. The detailed descriptions of the algorithms proposed in this paper are given in Chen (2000). In
section 3, the accuracy of the DEM derived and the GCPs generated automatically are verified. The geocoded SAR
images derived, as shown in section 4, are used to verify the proposed algorithms.
2 ALGORITHMS
2.1 Automatic Generation of DEMs from Stereo SAR Data
Generation of spatial data from a pair of stereo SAR images requires a geometric model and an image correlation
method. An ideal geometric model for extracting spatial data from a SAR image pair is expected to obtain high quality
results, to require a minimal number of GCPs and human interventions. The range/Doppler equations can be used as
observation equations in a least squares adjustment for solving ground points from ERS data, (Chen and Dowman,
1996). Refinement of the algorithm has been made using a weighting matrix to cope with inferior orbit data provided by
RADARSAT SAR imagery in order that the effects from the azimuth timing error on the results derived can be
minimized. GCPs are not required in the geometric model. Systematic correction can be made using only two GCPs
after space intersection. An area-based least squares correlation method, (Gruen, 1995), with a region-growing
approach, (Otto and Chau, 1989), is adopted to generate a dense coverage of elevation data set. This method has been
refined using a pyramidal harness to propagate control from the top to bottom image tier hence the effect of speckle in
SAR imagery can be reduced, (Denos, 1992). An optimized strategy of parameter determination for commanding the
pyramidal structure has been proposed to improve the level of automation, provided that the maximal parallax of the
image pair to be correlated can be measured or calculated using a coarse DEM, (Dowman and Chen, 1998).
2.2 Automatically Generated Control
In order to reduce manual operations, and then to stabilize the quality of GCPs, a SAR image product simulation
method is proposed to automatically provide ground control for radargrammetry. The proposed simulator is based on
the assumption of the perpendicular relationship between the resultant range and velocity vectors in SAR imagery. A
known DEM chip of a small area with varied terrain surface and the header/orbit data of a real SAR image are required.
Also, it is assumed that the structure of a real SAR image is well defined by the header/orbit information allowing
automatic detection of the extent of the simulated area without human intervention. The geometric considerations for a
SAR image simulator include the simulation of the image structure of a real image with no manually defined parameters
required, such as the range and azimuth time offsets for a sub-scene SAR image and incidence angles for individual
pixels. In terms of radiometry, a simple reflectivity model is applied to the entire simulated area without regard to
speckle or different ground cover. The reflectivity model is defined as the scalar product of a range vector from the
sensor to a DEM element and the outward surface normal of that DEM cell of interest. This simplification cannot be
avoided because the detailed ground truth data of an arbitrary real image, which affects the reflectance of a radar wave,
is not always available.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B1. Amsterdam 2000.
39
Pu-Huai Chen
2.2.1 SAR Image Simulation. The track of a sensor in 3-D space is a smooth curve that is continuous and
differentiable. There always exists a tangent passing through a point on the orbit and a normal plane perpendicular to
the tangent at that point. Physically, the position vector of a SAR sensor is changing along its orbit corresponding to its
azimuth time and the tangent at that point is equivalent to the sensor velocity vector. The normal plane passing through
that point represents a zero-Doppler plane with an equal range circle from the sensor to ground points. The concept of
the proposed SAR simulation method is based on the zero-Doppler imagery and on a scanning process from the first
azimuth line to the last one through the entire simulated area for detecting any DEM element encountered. Each normal
plane derived along the orbit track resembles the corresponding azimuth line of a simulated SAR image.
2.2.2 Real-Simulated Image Correlation. Four corner points of the known DEM chip are treated as GCPs whose
ground co-ordinates are already known, but the real image co-ordinates of the GCPs have to be derived using a realsimulated image correlation technique. The well-known normalized cross-correlation measure is adopted to search for
the best match of the corresponding position of the simulated image within the real image co-ordinate system. The
simulator also records the simulated range and azimuth time data of the GCPs based on the SAR imaging geometry
when the GCPs are detected in the process of simulation. The simulated range and azimuth time data of the GCPs can
be compared with those observations of the real image that gives the range and azimuth timing shifts and can be used to
provide control in geocoding.
2.3 Refinement of the DEM
Using the rigorous algorithm of space intersection and the pyramidal image correlation scheme with a region-growing
technique directed by an optimized parameter selection strategy, a raw DEM composed of irregularly distributed ground
points can be derived. Further effort for the automatic provision of ground control that improves the accuracy of the
GCPs used, ensures the quality of the radargrammetric results derived. The DEM obtained is, however, incomplete in
terms of practical applications, which demand spatial data without gaps or blunders. There are typical terrain induced
effects in a SAR image pair, for example, layover, leading to the formation of gaps in a raw DEM generated. These
effects decrease the quantity, and deteriorate the quality, of the results derived from image correlation routines. A
method of data fusion for the DEMs generated can be made to improve the DEM coverage without degrading the
overall accuracy, for example, filling the major gaps using the DEMs generated from different orbit directions. Since
the original elevation data is not changed, there is no effect on the quality of the DEM derived, but the coverage of that
DEM can be improved efficiently. Further improvements can be made using an interpolation method to fill minor gaps.
This makes a ‘complete’ DEM without gaps or voids. In addition, a filtering technique is employed to suppress noise
and to derive a practical DEM. Filtering DEMs can have a significant effect on the final results. A quality check for the
DEM derived is essential for each step of refinement.
2.4 SAR Image Geocoding
Geocoded SAR images have been used as an application example of the stereo-generated DEM to validate the
proposed algorithms. Thus, an algorithm with the minimal degradation of accuracy for the SAR image geocoding must
be considered. Precise terrain corrected geocoding of SAR imagery requires a known DEM, the imaging geometry/orbit
data and GCPs or tie points. Recent geocoding methods for ERS-1 data employ a map-to-image transformation based
on the range-Doppler equations, for instance, Dowman et al. (1993). The range-Doppler approach is performed only for
an anchor point grid (or a super-grid structure) constructed by regularly distributed DEM elements to reduce
computational load. All of the other DEM elements are then transformed into image co-ordinates by an efficient bilinear
interpolation (or other polynomial fit functions), according to the anchor point grid established. A direct approach
relating the SAR image to object space has been developed that ensures the minimal accuracy degradation in
geocoding. The proposed geocoding process resembles the principle of the proposed method of SAR image simulation.
Since no polynomial fit functions are used, the risks of possible errors caused by the polynomial fit functions can be
completely avoided. Absolute ground control for geocoding is usually provided using the manually collected GCPs
whose quality is not always stable. In order to get rid of any human interventions for collecting GCPs, the automatically
generated GCPs are used to correct the range and azimuth timing errors in the header/orbit files.
3 TEST DATA AND VERIFICATION OF THE DEM AND GCPS GENERATED
Under the RADARSAT Application Development and Research Opportunity (ADRO) Programme, four RADARSAT
images provided by the RADARSAT International (RSI) have been tested for generation of DEMs and geocoded
images. These images cover the Aix-Marseilles test area in South France with varied terrain utilization and spatial
40
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B1. Amsterdam 2000.
Pu-Huai Chen
characteristics. The variation of terrain surface ranges from 0m to 640m above mean sea level. The SAR images used in
experiments are grouped into two pairs according to orbit direction, as listed in Table 1.
Orbit
direction
Descending
pair
Ascending
pair
RADARSAT
beam mode
Standard-1
Standard-7
Standard-1
Standard-7
Acquisition
Full-scene size
Sub-scene for image correlation
date
(row/column)
(row/column)
1997-08-22
7901 / 9075
1024 / 1024
1998-03-13
7875 / 8937
1024 / 1024
1998-03-01
7876 / 8961
1024 / 1024
1997-08-08
7901 / 9139
1024 / 1024
Table 1. Four SAR images being simulated.
Alias
name
DS1
DS7b
AS7
AS1
Notice that the acquisition time of each image varies by more than six
months! Using the rigorous algorithm of space intersection and the
pyramidal image correlation scheme with a region-growing technique
directed by an optimized parameter selection strategy, two raw DEMs
were derived from the ascending and descending image pairs. Using
straightforward data addition, two raw DEM data files can be merged
together giving a new DEM with better coverage, but no quality
degradation. The DEM generated was interpolated into a 12.5m grid and
sectioned for the effective area of 10km × 10km, as shown in Figure 1.
The interpolated DEM can be refined further using a moving average
filter (radius=3 pixels) to suppress possible noise. The statistics of the
accuracy of the DEM derived from each step of refinement, as shown in
Table 2, were compared with a reference DEM of the test area, as shown
in Figure 2.
The reference DEM is of size 13.5km × 13.5 km (271 × 271 pixels) in a
Figure 1. The DEM generated and
50m grid with accuracy of 5m in elevation and in the French Lambert
refined (10km × 10km)
Conformal Conic Zone III map projection system. In order to provide
control in radargrammetry, two manually
selected GCPs could be used to provide
control. However, since a known DEM chip
of size 1km × 1km (21 × 21 pixels) is
available, as shown in Figure 2 (right), it
was used for SAR image simulation to
generate four GCPs automatically showing
that the systematic shift in the raw DEMs
can be eliminated as demonstrated in Table
2. As mentioned before, the simulated
image based on the known DEM chip has to
be correlated with the real image to provide
control. Examples of the simulated image
chips are shown in Figure 3 (right) with the
correlated real counterpart as shown in
Figure 3 (left). The size of each simulated
image chip varies according to the radar
Figure 2. Left: reference DEM (IGN ©, size: 13.5km × 13.5km in a
illumination and direction of the real orbit
50m grid,). Right: a DEM chip (1km × 1km) to be used in SAR
track, as shown in Table 3.
image simulation for providing control.
DEM Type
Original DEM
Refined DEM
DEM Processing
GCPs
Mean (m)
RMS (m)
(descending pair)
Manual
+1.4
23.9
(descending pair)
Automatic
+1.4
23.9
(ascending Pair)
Manual
23.4
10.2
(ascending Pair)
Automatic
23.6
+2.3
Automatic
Data addition
+1.8
23.8
Automatic
Interpolation
-1.595
21.63
Filtering (radius=3) Automatic
-1.596
20.74
Table 2. Statistics of the generated DEMs.
Min. (m)
-189
-189
-209
-217
-217
-192
-170
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B1. Amsterdam 2000.
Max. (m)
+223
+223
233
+219
+223
+138
+132
41
Pu-Huai Chen
The calculated range
and azimuth time of the
four GCPs generated
automatically
were
compared
with
the
observations of the real
(a)
(b)
SAR images giving
range and azimuth time
shifts as shown in Table
4. It can be seen that
there are significant
shifts (mean values in
Table 3) in the orbit
(c)
(d)
data, which differ from
–41m to +110m in range
Figure 3 Real (left)-simulated (right) image chips. (a) DS1 (93 × 121 pixels). (b) DS7b
and –151m to +61m in
(89 × 100 pixels). (c) AS1 (95 × 96 pixels). (d) AS7 (98 × 101 pixels)
azimuth direction in
terms of the product of the averaged sensor speed (7500m/sec) and the azimuth time shift. Also, it is obvious that the
RMS error of individual GCPs derived automatically is less than 0.4 pixel of a RADARSAT standard image. The range
and timing shifts will be considered and imported to the geocoding routine to correct orbit data errors.
Alias name
DS1
DS7b
AS1
Simulated image size (row/column)
93 / 121
89 / 100
95 / 96
Table 3. Statistics of the four automatically generated GCPs.
AS7
98 / 101
Alias name
DS1
DS7b
AS1
AS7
Range
Mean (m)
+94.2
-41.4
+8.3
+110.0
Difference
RMSE (m)
2.6
2.2
1.6
3.4
Azimuth
Mean (sec)
-0.00617
-0.02023
+0.01412
+0.00825
Timing
(m)
(-46.0)
(-150.7)
(+105.2)
(+61.5)
Difference
RMSE (sec)
0.00027
0.00008
0.00062
0.00017
(m)
(2.0)
(0.6)
(4.6)
(1.3)
Table 4. Statistics of the four automatically generated GCPs.
4 TEST OF GEOCODING
4.1 Geocoding Using the Reference DEM
An example of the geocoded images (DS1) derived using
the reference DEM is shown in Figure 4. The reference
DEM has to be interpolated into a 12.5m grid to cope with
the pixel spacing of the geocoded image required. Seven
check points were selected from the test area to evaluate
the accuracy of the geocoded image derived. Due to the
difficulty in identification of the image features in SAR
images, all of the check points measured are the corners of
buildings. The elevation of each object measured is not
known.
The co-ordinates in plan of each check point measured
from the geocoded image and from a 1:25,000 scale map
were compared, giving statistics in respect to each
geocoded image as in Table 4 and showing that the RMS
errors are below 21m in easting (E) or northing (N).
Figure 4. Geocoded image DS1 derived using the
Notice that the nominal resolution of a RADARSAT
reference DEM (11km × 11 km)
standard image is 25m. In addition, the elevation of each
object measured, which is liable to cause layover effect
and local displacement in plan, has not been corrected. The deviations derived show the resultant accuracy of the
reference DEM, the performance of the geocoding algorithm and the measuring error of a human operator. The
42
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B1. Amsterdam 2000.
Pu-Huai Chen
deviations in E or N of check points calculated and averaged from four geocoded images are also shown in Figure 5 to
demonstrate the error behavior. Generally speaking, there is no significant systematic shifts being shown. It proves that
the automatically generated GCPs do work reasonably. However, the elevation of each object measured is strongly
related to layover effect as can be seen from the difference of mean values shown in Table 4, particularly in the cases of
high-incidence-angle images, such as DS1 and AS1. Also, the human operator’s error cannot be underestimated due to
the difficulty in the process of manual map-image correlation. Tests of geocoding for the DEM generated and refined
are given in next section.
Alias Name
DS1
DS7b
AS1
AS7
Mean (m)
Component
E
N
E
N
E
N
E
N
E
N
Mean (m)
-32
+1
+15
+3
-19
+9
+2
+2
-8
+3
RMSE (m)
21
16
13
21
21
18
18
19
2-D (m)
26
25
28
26
Table 4. Positioning accuracy of the check points measured in respect to each geocoded image derived using the
reference DEM.
Check Point
Check Point
Deviation (m)
Deviation (m)
40
20
0
-20
-40
20
0
-20
-40
3119
3118
3117
3116
3115
3114
3113
3112
3111
3110
3109
3108
856
855
854
853
852
851
850
849
848
847
Easting (km)
40
Northing (km)
Figure 5. Deviations of the check points (in easting, left, and in northing, right) measured and averaged from the four
geocoded images derived using the reference DEM.
4.2 Geocoding Using the Derived and Refined DEM
Based on the same ground control as described in Section
4.1, and the RADARSAT SAR orbit data, geocoded
images were generated using the derived and refined
DEM, as shown in Figure 6. In terms of visual comparison
between Figure 4 and 6, the differences of both images are
concentrated in the hilly areas (the central and
southeastern part of the test area). This is because
information extraction in hilly areas, where layover effect
is significant, is incomplete, hence the DEM derived in
hilly areas is relatively less reliable.
The same seven check points were selected from the test
area to evaluate the accuracy of the geocoded image
derived using the generated and refined DEM. The coordinates in plan of each check point measured from the
geocoded image and from a 1:25,000 scale map were
compared giving statistics in respect to each geocoded
image as in Table 5 showing that the RMS errors are
below 24m in easting (E) or northing (N). Table 4 and 5
Figure 6. Geocoded image DS1 derived using the DEM
shows similar RMS error magnitudes in E or N. The mean
generated and refined ( Size: 11km × 11 km)
values of positional errors in terms of the geocoded images
derived using the generated DEM, as shown in Table 5, are slightly better than those of the geocoded images derived
using the reference DEM. It is probably because the reference DEM gives elevation at the ground level, but the
generated DEM is referred to the ‘rooftop’ level (of ground cover), not ground level. The deviations in E or N of check
points calculated and averaged from four geocoded images are also shown in Figure 7 to understand the error behavior.
Notice that Figure 5 and 7 demonstrate a similar trend of error curve. It means that the accuracy of the DEM generated
and refined is comparable with that of the reference DEM in flat-moderate areas.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B1. Amsterdam 2000.
43
Pu-Huai Chen
Alias Name
DS1
DS7b
AS1
AS7
Mean (m)
Component
E
N
E
N
E
N
E
N
E
N
Mean (m)
-16
+7
0
+5
-10
+7
-7
-4
-8
+4
RMSE (m)
15
10
14
24
20
13
20
20
2D (m)
18
28
24
28
Table 5. Positioning accuracy of the check points measured in respect to each geocoded image derived using the
generated DEM.
Check Point
Check Point
Deviation (m)
Deviation (m)
40
20
0
-20
0
-20
-40
3119
3118
3117
3116
3115
3114
3113
3112
3111
3110
3109
856
855
854
853
852
851
850
849
848
847
Easting (km)
20
3108
-40
40
Northing (km)
Figure 7. Deviations of the check points (in easting, above, and in northing, below) measured and averaged from the
four geocoded images derived using the generated DEM.
In order to give a close look at the
geocoded images, Figure 8 shows
the image chips extracted from the
geocoded images (DS1) derived
using the generated DEM (left),
those using the reference DEM
(central)
and
the
absolute
difference image of both images
(right). The darker the right image,
the greater the difference in grey
value. The size of each image chip
in Figure 8 is 2.5km × 2.5km. The
geocoded images in the top row are
sectioned within a hilly area. The
geocoded images in the bottom row
are extracted from a flat-moderate
area.
Obviously, the absolute difference
Figure 8. Geocoded image chips (DS1) in a hilly area (top row) and in a
image shows that image distortion
moderate-flat area (bottom row). (Left: the geocoded image using the
in the hilly areas is evident (images
reference DEM. Central: the geocoded image using the DEM derived. Right:
in the upper row), but not in the
the difference image of the left and the central images. Size: 2.5km ×2.5km.
moderate-flat areas. Notice that the
The darker the right image, the greater the difference in grey values.)
roads (the dark lines) of the left and
middle geocoded images in Figure
8 cannot be identified in the absolute difference image, i.e., the geometrically rectified locations of the roads in both
images are overlapping each other (see the right images). Most of the bright objects (buildings) in the geocoded images
of both rows are also overlapping each other, the high brightness spots showing correct geocoding. Particularly, the
location of a train station in the images in the bottom row has been geometrically rectified correctly giving a bright
linear feature. It proves that the quality of the generated DEM is suitable for geocoding in flat-moderate areas.
However, the layover areas as shown at the bottom part of the geocoded images in the upper row cannot be rectified
completely giving the dark areas as shown at the bottom of the upper-right image.
44
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B1. Amsterdam 2000.
Pu-Huai Chen
5 CONCLUSIONS
According to the results shown above, the rigidity and robustness of the proposed algorithms have been demonstrated.
A higher level of automation for extracting spatial information and providing control is achieved and there is no
manually selected GCPs required. The quality of the automatically generated GCPs assures that reliable ground control
is provided. Geocoding employing the DEM generated using the stereo-SAR method is practical in moderate relief
areas, provided that layover effects on SAR imagery are minimal.
ACKNOWLEDGMENTS
The authors are grateful to RADARSAT International (RSI) for providing the RADARSAT SAR images under the
Application Development and Research Opportunity (ADRO) Programme. P.-H. Chen is grateful to the Government of
the Republic of China and the Chung Cheng Institute of Technique for the full sponsorship of overseas Ph.D. study at
University College London in U.K.
REFERENCES
Chen, P.-H., 2000. Extraction of Spatial Information from Stereoscopic SAR Images. Ph.D. thesis submitted to
University of London. 237 pages.
Chen, P.-H. and Dowman, I. J., 1996. Space intersection from ERS-1 synthetic aperture radar images. Photogrammetric
Record, 15 (88):561-573.
Clark, C., 1991. Geocoding and Stereoscopy of Synthetic Aperture Radar Images. Ph.D. Thesis, University of London.
223 pages.
CSA, 1995. RADARSAT Canadian Data Processing Facility Product Specification. Canadian Space Agency. CDRL
No.IS-3, MacDonald-Dettwiler, Canada.
Denos, M., 1992. A pyramidal scheme for stereo matching SIR-B imagery. International Journal of Remote Sensing,
13(2):387-392.
Dowman, I.J. and Chen, P.-H., 1998. A rigorous stereo method for RADARSAT data. RADARSAT ADRO Final
Symposium, Montreal. CD-ROM.
Dowman, I.J, Laycock, J. and Whalley, J., 1993. Geocoding in the UK. In SAR Geocoding: Data and Systems. Edited
by Schreier, G. Wichmann Verlag, Heidelberg, pp:373-387.
Guindon, B., 1995. Performance evaluation of real-simulated image matching techniques in the acquisition of ground
control for ERS-1 image geocoding. ISPRS Journal of Photogrammetry and Remote Sensing, 50(1): 2-11.
JPL, 2000. Shuttle Radar Topography Mission (SRTM). Jet Propulsion Laboratory.
http://www.jpl.nasa.gov/srtm (17 March 2000).
Kowk, R., Curlander, J.C. and Pang, S.S., 1990. An automated system for mosaicking spaceborne SAR imagery.
International Journal of Remote Sensing, 11(2):209-223.
Leberl, F.W., 1990. Radargrammetric Image Processing. Artech House Inc., Norwood, U.S.A. 595 pages.
Otto, G.P. and Chau, T.K.W., 1989. Region-growing algorithm for matching of terrain images. Image Vision
Computing, 7(2):83-94.
Twu, Z.-G., 1996. Automatic Height Extraction from Stereoscopic SAR Imagery. Ph.D. Thesis, University of London.
222 pages.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B1. Amsterdam 2000.
45